I would like to estimate reliability of the MNE activations on a group level. dSPM, that give us F values, can be performed in a single subjects, but what is about GROUP dSPM?
I have seen the presentation by Daniel Goldenholz, where he speculates on possibility of the ?group dSPM? and I have seen the post by Marina Papoutsi (from February 18, 2009), where she tried to implement this analysis. However, there was no news on this topic since then. Does anybody use the ?group dSPM? and/or has any practical hints on that?
there is one way that would consist is keeping track on the noise normalization
used for each subject dSPM, then morph the noise normalization values too
and then do the maths to average MNE estimates and divide by the mean
noise normalization.
Unfortunately there is build in way to do this currently...
Hi Elena,
If you are using a fixed orientation solution, then the dSPMs are z
scores and can be combined in group studies using t-stats etc..
If you are using a free orientation solution, then depending on how you
combine the compnents you end up with an F or a chi2.. Atleast as a first
pass, you could compute group stats using a transform such as that goes
from chi2 (with 3 dof) to Z such as:
y = (x/3)^(1/3)
y is then approximately distributed normally with mean (1 - 2/27)and
variance (2/27)
See: Wilson, E.B.; Hilferty, M.M. (1931) "The distribution of
chi-squared". Proceedings of the National Academy of Sciences, Washington,
17, 684?688.
We have usually dSPM scores non-parametrically (i.e using permutation type
tests) at the group level which might be the cleanest way of doing it but
very slow.
See: http://www.ncbi.nlm.nih.gov/pubmed/11747097
MNE assembles orientation with sqrt(x^2 + y^2 + z^2)
The problem that can happen is that since you're likely to use loose
orientation the variance of x, y and z component will be different
but also the number of trials for each subject will vary causing
amplitude differences in dSPM values.
@hari : do you agree? From your experience does your chi2->T transformation
somehow reduces the problem?
If you combine using A = x^2 + y^2 + z^2 (i.e without the square root or
square what MNE gives you) and if x,y,z are *after noise normalization*,
then it is reasonable to assume that A is chi2 as long as the noise
covariance was computed using a large number of points, I wouldn't be
concerned about variances being different since each has approximately
variance 1 in the null...
The maps we have gotten from running long permutation tests at the group
level and then thresholding using this transformation look very similar.
So our current practice is to use this transform and run the usual
parametric analyses and invest time in running permutations as
confirmation once we see something we like.
?If you combine using A = x^2 + y^2 + z^2 (i.e without the square root or
square what MNE gives you) and if x,y,z are *after noise normalization*,
then it is reasonable to assume that A is chi2 as long as the noise
covariance was computed using a large number of points, I wouldn't be
concerned about variances being different since each has approximately
variance 1 in the null...
the noise normalization is the same for x, y and z and as you regularize
more the tangential components than the radial I don't think the
variance will be
the same even when you apply dSPM to noise. But I should check as it's
just an intuition.
The maps we have gotten from running long permutation tests at the group
level and then thresholding using this transformation look very similar.
So our current practice is to use this transform and run the usual
parametric analyses and invest time in running permutations as
confirmation once we see something we like.
if the noise normalization is different for the 3 orientations you
cancel the effect
of the fixed or loose orientation and end up with solutions that look very much
like free orientation solutions.
Do you think it is possible to estimate significance of the group mean dSPM values?
In case of loose constrains the dSPM values come from F distribution with dof 3 and Npoints x 3 (Dale, 2000). One can e.g. calculate distribution of the means of N (N=number of subjects) randomly generated F values and empirically assess what value would correspond to e.g. p=0.05.
Hi Elena,
If you have roughly similar #trials across subjects for noise-cov
estimation and averaging, that is very reasonable if you are looking
at an ROI.. However, in order to deal with the multiple comparison
problem across vertices, you'll have to do some kind of
bootstrapping/permutations to determine what the p = 0.05 threshold at
the whole brain level is... If your effects are strong, you could be
conservative and do a vertewise bonferroni correction or FDR..